Can Foundation Models Label Data Like Humans? Hugging Face Analysis

Hugging Face researchers investigated whether foundation models, specifically GPT-4, can reliably label data and evaluate model preferences as a proxy for human judgment. The study reveals that while GPT-4 can mirror some human rankings, it suffers from significant positional bias and a preference for verbosity over conciseness, making it an imperfect substitute for human annotators.

LLM Preference Evaluation Methodology

Hugging Face expanded the Open LLM Leaderboard to include a controlled study comparing human labels (via Scale AI) and GPT-4 evaluations. The researchers used a held-out set of 327 high-quality, human-written prompts across categories such as generation, brainstorming, question answering, summarization, commonsense, and coding.

Four open-source models were evaluated: Vicuna-13B, Koala-13B, OpenAssistant-12B, and Dolly-12B. Raters (both human and GPT-4) used a Likert scale from 1 to 8 to indicate preference between two model completions, where 1 represents a strong preference for the first model and 8 for the second.

Human vs. GPT-4 Elo Rankings

Both human and GPT-4 evaluators produced similar relative rankings of the models, but the margins differed. In both sets of results, Vicuna-13B consistently ranked highest, followed by Koala-13B, OpenAssistant-12B, and Dolly-12B.

Human Elo Results (Median)

Model Elo Ranking (No Ties) Elo Ranking (With Ties)
Vicuna-13B 1140 1130
Koala-13B 1073 1061
Oasst-12B 986 988
Dolly-12B 802 820

GPT-4 Elo Results (Median)

Model Elo Ranking (No Ties) Elo Ranking (With Ties)
Vicuna-13B 1134 1114
Koala-13B 1082 1082
Oasst-12B 972 973
Dolly-12B 812 831

Critical Biases in LLM-Based Evaluation

The study identified several systemic biases that affect the reliability of GPT-4 as an evaluator:

Positional Bias

GPT-4 exhibits a strong positional bias, frequently favoring the first response presented in a pairwise comparison (rating it as "1" on the 1-8 scale), regardless of the actual quality of the content.\n

Verbosity and Style Bias

GPT-4 tends to prefer longer, more detailed responses over concise, correct ones. This is evidenced by examples where GPT-4 rated a verbose model response as "greatly better" than a brief, accurate human response. This suggests a "style benefit" for models trained on outputs from other large language models (e.g., Vicuna), which the researchers describe as "unintentional doping."

Human Demonstration Paradox

When human-written demonstrations were included in the evaluation, GPT-4 ranked them lower than several open-source models (e.g., Vicuna and Koala), further highlighting the model's preference for LLM-style verbosity over human-style conciseness.

Correlation and Task Performance

The correlation between GPT-4 and human labels varies significantly by task type. GPT-4 is more aligned with humans in high-entropy, creative tasks than in low-entropy, factual or technical tasks.

Category Correlation: GPT-4 to Human Labels
Brainstorm 0.60
Creative generation 0.55
Commonsense reasoning 0.46
Question answering 0.44
Summarization 0.40
Natural language to code 0.33

Key Takeaways for LLM Evaluation

  • Positional Bias is Persistent: Asking GPT-4 to "de-bias" itself often simply flips the bias in the opposite direction rather than eliminating it.
  • Verbosity $\neq$ Quality: GPT-4 rewards token quantity, which can lead to misleading rankings where verbose but less useful models outperform concise and accurate ones.
  • Task-Dependent Reliability: LLM-based evaluation is most reliable for brainstorming and creative generation and least reliable for coding tasks.
  • Formatting Limitations: The researchers noted that models like ChatGPT often struggle to return answers in the correct format for Likert scales, indicating that formatting control is a prerequisite for becoming a useful evaluation tool.

Sources